app.py

import gradio as gr
import torch
from transformers import (
    BertTokenizer,
    XLNetTokenizer,
    GPT2Tokenizer,
    AutoModelForSequenceClassification
)

# Model repositories on Hugging Face Hub
model_repos = {
    "BERT": "sk23aib/emotion-bert",
    "XLNet": "sk23aib/emotion-xlnet",
    "GPT-2": "sk23aib/emotion-gpt2"
}

# Emotion labels (must match model training order)
emotion_labels = [
    "anger", "boredom", "empty", "enthusiasm", "fun", "happiness", "hate",
    "love", "neutral", "relief", "sadness", "surprise", "worry"
]

# Load models and tokenizers
loaded_models = {}

# BERT
bert_tokenizer = BertTokenizer.from_pretrained(model_repos["BERT"])
bert_model = AutoModelForSequenceClassification.from_pretrained(model_repos["BERT"])
bert_model.eval()
loaded_models["BERT"] = {"tokenizer": bert_tokenizer, "model": bert_model}

# XLNet
xlnet_tokenizer = XLNetTokenizer.from_pretrained(model_repos["XLNet"])
xlnet_model = AutoModelForSequenceClassification.from_pretrained(model_repos["XLNet"])
xlnet_model.eval()
loaded_models["XLNet"] = {"tokenizer": xlnet_tokenizer, "model": xlnet_model}

# GPT-2
gpt2_tokenizer = GPT2Tokenizer.from_pretrained(model_repos["GPT-2"], padding_side="left")
gpt2_tokenizer.pad_token = gpt2_tokenizer.eos_token  # Required for GPT-2
gpt2_model = AutoModelForSequenceClassification.from_pretrained(model_repos["GPT-2"])
gpt2_model.config.pad_token_id = gpt2_tokenizer.pad_token_id
gpt2_model.eval()
loaded_models["GPT-2"] = {"tokenizer": gpt2_tokenizer, "model": gpt2_model}

# Inference function – return top emotion + probability
def predict_emotions(text):
    output_lines = []
    with torch.no_grad():
        for model_name, components in loaded_models.items():
            tokenizer = components["tokenizer"]
            model = components["model"]
            inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True)
            logits = model(**inputs).logits
            probs = torch.nn.functional.softmax(logits, dim=1)[0]
            top_idx = torch.argmax(probs).item()
            top_emotion = emotion_labels[top_idx]
            top_confidence = round(float(probs[top_idx]), 4)
            output_lines.append(f"{model_name}: {top_emotion} ({top_confidence})")
    return "\n".join(output_lines)

# Gradio Interface
interface = gr.Interface(
    fn=predict_emotions,
    inputs=gr.Textbox(lines=3, placeholder="Type a sentence to analyze..."),
    outputs=gr.Textbox(label="Top Emotion by Model"),
    title="Multi-Model Emotion Classifier",
    description="See which emotion is predicted by BERT, XLNet, and GPT-2, along with their confidence."
)

interface.launch()